Utillisation dessignaturesdans lesessaiscliniques en oncologie
S.Michiels
CESP,INSERMU1018,UniversityParis-Saclay,UniversityParis-Sud,Villejuif,FranceServicedeBiostatistiqueetd'Épidémiologie,GustaveRoussy,Villejuif,France
Today’s talk
• Prognostic gene signatures
• Predictive gene signatures
• Prospective clinical trials
2
Lessons from early breast cancer: prognostic signatures
Available gene signatures for a price of 400-4175$…• IHC4: 4 genes
• Oncotype Dx: 16 cancer genes
• PAM50 (ROR): 50 genes
• Mammaprint Dx: 70 genes
• Endopredict: 8 cancer genes
• Mapquant Dx (GGI) : 97 genes
• …
European consensus panel (Azim, Michiels Ann Onc 2013): no robust evidence of clinical utility … 3
Pooled gene expression analysis of neoadjuvant chemotherapy trials in breast
cancer: flow chart
Ignatiadis et al JCO 2012 4
OR: odds ratio for a 1-unit changein gene module, adjustedfor clinicopathological factors
FDR: false discovery rate
GGIGene70CIN70Stroma1Stroma2Immune1Immune2RASMAPKPTENAKTmTORPIK3CAIGF1SRCMYCE2F3BetaCatenin
OR1.592.111.470.650.661.761.490.770.811.710.84
10.791.021.131.671.12
(0.91,2.81)(1.12,4.03)(0.88,2.47)(0.38,1.08)(0.4,1.08)
(1.13,2.76)(0.96,2.31)(0.5,1.19)
(0.47,1.38)(1.04,2.85)(0.54,1.3)(0.55,1.8)
(0.46,1.34)(0.62,1.65)(0.73,1.75)(1.06,2.67)(0.69,1.82)
1.0E 012.2E 021.4E 019.7E 029.9E 021.3E 027.4E 022.4E 014.3E 013.7E 024.3E 019.9E 013.8E 019.5E 015.7E 012.9E 026.4E 01
2.2E 011.6E 012.7E 012.2E 012.2E 011.6E 012.2E 014.1E 015.7E 011.6E 015.7E 019.9E 015.7E 019.9E 016.9E 011.6E 017.3E 01
GGIGene70CIN70Stroma1Stroma2Immune1Immune2RASMAPKPTENAKTmTORPIK3CAIGF1SRCMYCE2F3BetaCatenin
OR0.860.850.760.650.467.135.010.341.041.130.391.3
0.791.740.770.651.68
(0.25,2.72)(0.24,2.8)
(0.22,2.41)(0.25,1.62)(0.17,1.19)
(2.38,25.51)(1.77,16.35)
(0.09,1.12)(0.37,2.84)(0.38,3.31)(0.14,0.99)(0.47,3.69)(0.3,1.94)
(0.73,4.33)(0.31,1.82)(0.21,1.97)(0.68,4.35)
8.0E 018.0E 016.4E 013.5E 011.2E 011.0E 034.1E 038.7E 029.4E 018.2E 015.4E 026.2E 016.2E 012.2E 015.6E 014.5E 012.7E 01
8.7E 018.7E 018.4E 017.5E 014.0E 011.8E 023.5E 023.7E 019.4E 018.7E 013.1E 018.4E 018.4E 016.1E 018.4E 018.4E 016.5E 01
GGIGene70CIN70Stroma1Stroma2Immune1Immune2RASMAPKPTENAKTmTORPIK3CAIGF1SRCMYCE2F3BetaCatenin
OR3.193.433.381.051.33
1.41.711.810.863.061.780.841.780.631.072.560.43
(1.26,8.72)(1.25,9.66)(1.31,9.35)(0.45,2.42)
(0.57,3.1)(0.59,3.15)
(0.7,4.17)(0.64,5)
(0.32,2.16)(1.1,8.94)
(0.66,5.02)(0.38,1.83)(0.65,4.96)(0.23,1.63)(0.41,2.73)(1.14,5.97)(0.17,1.09)
1.8E 021.8E 021.4E 029.1E 015.1E 014.3E 012.4E 012.6E 017.5E 013.5E 022.6E 016.6E 012.6E 013.5E 018.9E 012.5E 027.5E 02
1.0E 011.0E 011.0E 019.1E 016.7E 016.1E 014.4E 014.4E 018.5E 011.2E 014.4E 018.0E 014.4E 015.4E 019.1E 011.1E 012.1E 01
ER /HER2
HER2+ ER+/HER2
(394 pts, 120 pCR)
(116 pts, 42 pCR) (335 pts, 27 pCR)
0.25 1 5 10 20
0.25 1 5 10 20 0.25 1 5 10 20
95% CI P FDR
95% CI P FDR 95% CI P FDR
Figure 4
Odds Ratio
Odds Ratio Odds Ratio
GGIGene70CIN70Stroma1Stroma2Immune1Immune2RASMAPKPTENAKTmTORPIK3CAIGF1SRCMYCE2F3BetaCatenin
OR1.7
2.021.610.730.741.921.780.820.851.750.841.010.971.02
1.11.6
0.98
(1.12,2.6)(1.29,3.2)
(1.08,2.42)(0.49,1.06)(0.5,1.07)
(1.36,2.73)(1.25,2.53)(0.57,1.18)(0.56,1.27)(1.18,2.62)(0.59,1.19)(0.67,1.53)(0.65,1.45)(0.71,1.47)(0.78,1.56)(1.12,2.3)
(0.68,1.43)
1.3E 022.4E 032.1E 021.0E 011.1E 012.2E 041.3E 033.0E 014.2E 015.8E 033.2E 019.5E 018.9E 019.1E 015.8E 011.1E 029.4E 01
3.7E 021.3E 025.1E 022.1E 012.1E 013.7E 031.1E 024.9E 016.0E 012.5E 024.9E 019.5E 019.5E 019.5E 017.6E 013.7E 029.5E 01
(845 pts, 189 pCR)
95% CI P FDR
ALL
Odds Ratio
0.25 1 5 10 20
A B
C D
Association of gene modules with pathological complete response after
anthracyclines, beyond clinicopathological factors
Ignatiadis et al JCO 2012
Randomized clinical trials and absolute risk
• Absolute risk reductions in large trials are still the best guide
• Absolute gain from a 33% breast cancer mortality reduction from chemotherapy(chemo) in hormone receptor-positive pts depends on the absolute risks with endocrine therapy alone
Rothwell Lancet 2005, Michiels, Rotolo 2015
Estimated absolute 10-year breast cancer-specific survival
No chemoAbsolute gain from chemo
97% +1%92% +3%85% +5%
We cannot necessarily, perhaps very rarely, pass from (the result of a clinical trial) to stating exactly what effect the treatment will have on a particular patient
Austin Bradford Hill, 1952
6
Lasso penalty in Cox model
• Popular approach to perform variable selection and develop a classfifer with high-dimensional data
• Estimation of tuning parameter (λ) usually via the cross-validated loglikelihood cvl : sum across K folds of the difference in penalized log-likelihood between the full sample and the training samples.
Verweij and van Houwelingen 1993
Drawback: Too many false positives selected• Theoretical results Meinshausen and Bühlmann 2006
• Empirical results Benner et al. 2010; Roberts and Nowak 20147
Objectives
- Selection of fewer biomarkers (i.e. λ ↗)- Trade-off between the goodness of fit (small λ) and parsimony (large λ)
Proposed penalty:
λ0 : the smallest λ for which no biomarker is included in the model pλ : number of non-null regression parameters for a given λ
Our empirical extension: lasso-pcvl
Ternès,Rotolo,MichielsStatMed2016
λ" = argmax) 𝑐𝑣𝑙(𝜆) − 𝑝𝑒𝑛(𝜆)
8
Evaluation of 11 alternatives
Penalization methods of the form 𝑝𝑒𝑛 𝜆 = 𝜃)𝑓)(𝑝))
Variants of the lasso penalty• One standard error rule (Friedman 2010)• Adaptive lasso (Zou 2006, Zhang et Lu 2007)• Percentile lasso (Roberts and Nowak 2014)• Stability selection (Meinshausen and Bühlmann 2010)
λ" = argmax) 𝑐𝑣𝑙(𝜆) − 𝑝𝑒𝑛(𝜆)
Estimationofthetuningparameter
9
Simulation studyData generation• Exponential survival times• n: 100, 500, 1000 patients• p: 10, 100, 1000 biomarkers• q: 0 (null), 1 and > 1 (alternative scenarios)• Equal effect size: β = log(0.8) (and varying effect sizes)• Autoregressive correlation between biomarkersEvaluation criteria
• FDR, FNR• Concordance index
Ternès,Rotolo,MichielsStatMed2016 10
Activebiomarkers (+)
Inactivebiomarkers (−)
Selectedbiomarkers
(+)
True Positive(TP)
False Positive(FP)
FalseDiscoveryRate
FP /(TP +FP)Unselectedbiomarkers
(−)
FalseNegative(FN)
TrueNegative(TN)
FalseNon-discoveryRate
FN/(TN+FN)Sensitivity
=TP/(TP +FN)Specificity
=TN/(TN+FP)
FalseNegativeRate
=FN/(TP+FN)
False PositiveRate
=FP/(TN+FP)
Null scenario (q = 0)
Ternès,Rotolo,MichielsStatMed2016 11
Alternative scenario (q > 1)
Ternès,Rotolo,MichielsStatMed2016 12
lasso-cvllasso-pcvl
lasso-cvl(1se)lasso-AIClasso-RIC
adaptivelassostabilityselection
2actives 10actives 20actives
Breast cancer example
Publicly available data from n = 523 patients with early breast cancer
Aim: Identify prognostic genes for distant recurrence-free survival
Ternès,Rotolo,MichielsStatMed2016
14
Standardized expression values of p = 1703 genes
Gene Signature DataBase⇒assess whether genes where also included in “known” curated breast cancer gene signatures
GATA3 also highly discussed in the biomedical literature
Median:22Range:14– 53
Median:16Range:6– 31
Wilcoxon P-value=0.02
Breast cancer example
Ternès,Rotolo,MichielsStatMed2016
0.5 0.6 0.8 1 1.3 1.6 2
Gemini/Libra
Other
Aspirin better Placebo better
ISIS-2 trial: effects of aspirin vs control on vascular death in 17,187 patients with acute myocardial infarction
Interaction p-value p=0.002
Odds ratio & 95% CIAstrologicalbirth sign
Whom to treat: subgroups?
Peto et al Lancet 198815
Start of the randomized clinical trial…
TheEconomist,200716
Randomize-all design
Buyse, Michiels et al, Expert Rev Mol Diag 2011
R
Biomarker -
Biomarker
Biomarker +Std
Exp Biomarker -
Biomarker
Biomarker +R:Randomization;Std:standardarm;Exp:experimental arm Testmanybiomarkerxtreatmentinteractions!
17
18
Multiple interaction testing throughpermutation in randomized trials
• Control of family-wise Type I error• Treatment-by-biomarker interactions in a Weibull model • 5 tests, using single-biomarker statistics and composite statistics
• Composite Difference statistic: |g1-g0|g0 = “concordance probability” that a randomly chosen control patient with a scaled biomarker score at 0 will outlive a treated patient with same score
• Permutation rearranges the patients within a treatment arm• In a strict sense not a permutation test of interaction but made
sensitive to detect interactions under the alternative…• Simulation study illustrating Type I error control and power when
there are ≥1 true treatment-effect modifiers and different correlation patterns
Michiels,Potthoff,GeorgeStatMed2011
Stepwise strategy for high-dimensionaldata in a randomized clinical trial (RCT)
• Step 1: Perform a global interaction test for control of the global Type I error at a prespecified α level (e.g. 5%)
• Step 2: Only when global test is significant, developclassifier in a survival model with interaction effects– Use 10-fold crossvalidation to estimate treatment
effects in composite biomarker score defined subsets– Applying survival model on full RCT data: indication
classifier for future patients
Michiels, Rotolo 201519
Step I: Global interaction test by permutation
Michiels, Potthoff, George Stat Med 2011; Michiels, Ternes, Rotolo (in revision)20
Step II: Develop a composite biomarker classifier through 10-fold crossvalidation
21Michiels, Ternès, Rotolo (in revision)
French breast RCT example• Tissue-array from two French breast cancer RCTs of
adjuvant chemotherapy with long term follow-up (798 pts)• 11 biomarkers (ER,PR, HER2, EGFR, p53, p27…)• Disease-free survival, 320 events
OverallHazardRatio:0.78[0.62–0.97]
0.0
0.2
0.4
0.6
0.8
1.0
Dis
ease
-free
sur
viva
l
No chemo Chemo
0 2 4 6 8 10Years
No. At RiskNo chemo 389 335 286 237 195 148
Chemo 409 368 321 266 223 161
ChemotherapyControl
22Michiels, Rotolo 2015
LowHigh
0.25 0.5 0.75 1 1.25 1.5
Hazard Ratio
French breast RCTs exampleStep I: Global test, 2000 permutations
• Composite difference test p= 0.01
Step II: Crossvalidated treatment effects according to composite biomarker score (cut-off at 0.5)
23Michiels, Rotolo 2015
Chemo better Control better
Lasso for many biomarker by treatment interactions…
24Ternès N,Rotolo F,HeinzeG,Michiels S(inrevision)
Thefullbiomarker-by-treatmentinteractionmodel T X1 X2 X3 TX1 TX2 TX3
Modifiedcovariates TX1 TX2 TX3
PenalizedregressionFull-lassoAdaptivelasso(sw:specificweights)Adaptivelasso(gw:groupedweights)Group-lassoRidge+lasso
T X1 X2 X3 TX1 TX2 TX3
T X1 X2 X3 TX1 TX2 TX3
T X1 X2 X3 TX1 TX2 TX3
T X1 X2 X3TX1 TX2 TX3
T TX1 TX2 TX3𝝋:;<<
DimensionreductionPCA+lassoPLS+lasso
T Z1 Z2 TX1 TX2 TX3
T Z1 TX1 TX2 TX3
Simulation study
Generation of p = 500 unit-variance Gaussian biomarkers for n = 500 patients (250 per arm), exponential survival
ALTERN
ATIVE
SCEN
ARIOS
Std Exp
Treatment
Clinicaloutcome
B+
B-
Std Exp
Treatment
Clinicaloutcome
B+
B-
Std Exp
Treatment
Clinicaloutcome
B+B-
±
(1) (10) (10)
Std Exp
Treatment
Clinicaloutcome
B+B-
Std Exp
Treatment
Clinicaloutcome B+
B-
Std Exp
Treatment
Clinicaloutcome
B+
B-NULL
SCEN
ARIOS (10)
Ternès N,Rotolo F,HeinzeG,Michiels S(inrevision)
HR=0.5
Simulation study: null scenarios
Ternès N,Rotolo F,HeinzeG,Michiels S(inrevision)
Proportion of models selecting at least one biomarker-by-treatment interaction among 250 replications è type-I error or FDR
Simulation study: alternative scenarios
Ternès N,Rotolo F,HeinzeG,Michiels S(inrevision)
Publicly available data in breast cancer patients (GEO)
Clinical datan = 614 patientsArm A: anthracycline only (n = 107)Arm AT: anthracycline + taxane-based
chemotherapy (n =507)
Standardized expression datap = 1689 genes
Aim: Identify treatment-modifiers for taxane in tumors biopsies from breast cancer
Breast cancer example
Ternès N,Rotolo F,HeinzeG,Michiels S(inrevision)
Trainingset(315patients) Validationset(299patients)
IFIH1 gene already known in the biomedical literature• Associated with recurrence in non-responders to taxane-based
chemotherapy in early breast cancer• Included in two patents that aim at predicting the benefit of taxanes in cancer
lasso
alasso
modifiedcovariates
ridge+lassoPCA+lasso
OAS1ARMCX2
MET
ITGB4
SGCB
AREG F12
NDRG2
AKAP9
GPD1LDDX17
ARMCX1
PDLIM4
SGK1
MTFR1
REN
HSPA1A SLC39A6
ME1
H1FX SOX11
SCGB2A2
CYP2B6DEFB1
KRT81EVL
CHPT1
IFIH1
Ternès N,Rotolo F,HeinzeG,Michiels S(inrevision)
Breast cancer example
ClinicalRisk
“DISCORDANT RISK” DESIGN
Low riskSignature
High riskLow risk
High risk
Low riskSignature
High risk
R
ChT
No ChT
Buyse, Michiels et al, Exp Rev Mol Diag 2011; ChT: chemotherapy
“INTERMEDIATE RISK” DESIGN
All patients R
Low risk
Signature
High risk
Intermediaterisk
No ChT
ChT
BENEFITS OF LARGE RANDOMIZED TRIALS FOR THE VALIDATION OF GENE SIGNATURES
• Tumor samples will be collected prospectively • Patients will be selected identically • Patients will receive controlled treatments• Randomization will make it possible to investigate the
predictive value of signatures• Micro-arrays and pathological reviews will be done
centrally• Large numbers will make the results reliable
R2:1
ArmA:Targeted therapy
ArmB:MaintenanceArm
N=650
N=230AZD2014 AZD4547 AZD5363
AZD8931 SelumetinibVandetanib
SAFIR02 Lung : STUDY DESIGN
NSCLCmetastatic orlocaly advanced
NoEGFRmutationNoALKtranslocation
Cis-platinumBasedregimen4cycles
Progression ortoxicity
PRorSDTargetablemolecularalteration
YES
NO
R2:1
MEDI4736
N=180
Pemetrexed Erlotinib
MaintenanceArm
Pemetrexed Erlotinib
Biopsymetastatic site:Next generation
sequencing PGMIontorrent35genesArray CGHAgilent
PI’s:JC.Soria,F.BarlesiOpen-label,multicentric phaseII
Clinicaltrials.gov:NCT02117167
SWOG lung master protocol
https://clinicaltrials.gov/ct2/show/NCT02154490
Estimated Enrollment:10000
Main references
• Michiels S,Potthoff RF, GeorgeSL.Multipletestingoftreatmentmodifyingbiomarkersinarandomizedclinicaltrialwithasurvivalendpoint,StatMed2011;30:1502-1518.
• Michiels S,Rotolo F.Evaluationofclinicalutilityandvalidationofgenesignaturesinclinicaltrials,inDesignandAnalysisofClinicalTrialsforPredictiveMedicine,CRCPress(eds MatsuiS.,Buyse M.,SimonR)2015.
• Ternès N,Rotolo F,Michiels S.Empiricalextensionsofthelassopenaltytoreducethefalsediscoveryrateinhigh-dimensionalCoxregressionmodels.StatMed2016;35(15):2561-73
• Ternès N,Rotolo F,Heinze GandMichiels S.Predictionoftreatmentbenefitinhigh-dimensionalcoxmodelsviagenesignaturesinrandomizedclinicaltrials.Trials 2015;16(Suppl 2),O86.